LUMAARCResearchAboutDevsChangelog
Decomps ↗NBA DECOMPS · COMING SOON · NBADECOMPS.COM_

LUMA Basketball Research · Model

LOKI predicted ARC

The best estimate of a player’s impact per 100 possessions for a given season, using only games played before the release date.

The 2026–27 board →
One-pager · Full paper coming soonARC one-pager →LOKI board →ARC board →Research →

Built on ARC, LOKI is a predictive model that estimates a player’s impact on both ends of the floor, accounting for aging, time decay and team context.

LOKI is an amalgamation of the ideas behind its base model, ARC (Adjusted Regularized Contribution), above all the philosophy of possession scarcity.

Rather than fitting box-score metrics to RAPM, LOKI takes the expected value of each box-score action within play-by-play, over millions of possessions, and blends it with pure RAPM inside the solve.

To make it predictive, LOKI models covariant subskills and aging, among other strategies within the solve, to project player talent. It adds the predictive signal from 6.5 billion tests of how more than 100 individual skills grow and regress over time.

In keeping with the nature of ARC, LOKI stays decomposable: an exact tree of parts rather than an opaque number, part of LUMA’s broader decomposition work across more than 100 channels. Tested out of sample, LOKI matches or beats DARKO and LEBRON at predicting seasons and games.

Preseason top 102026–27
  1. 1Shai Gilgeous-AlexanderOKC+12.6
  2. 2Nikola JokićDEN+10.9
  3. 3Victor WembanyamaSAS+10.3
  4. 4Giannis AntetokounmpoMIA+8.7
  5. 5Kawhi LeonardTOR+8.6
  6. 6Luka DončićLAL+7.2
  7. 7Chet HolmgrenOKC+6.4
  8. 8Joel EmbiidPHI+6.0
  9. 9Franz WagnerORL+5.7
  10. 10Donovan MitchellCLE+5.5
2007–082026–27

Reading LOKI

A player with a LOKI of 0 can be understood to have a neutral effect on their team’s expected point valuation per 100 possessions. This is because, intuitively, minutes are going to better players. The average possession is mostly defined by rotation players. Thus LOKI is functionally in units of points above the rotation player. The average LOKI of all 747 players on the 2026–27 board is −1.6.

-8-6-4-20+2+4+6+8+10+120 = average possessionTop 10#1–10+7.6Stars#11–30+4.0Starters#31–150+1.7Rotation#151–300−0.3Bench#301–450−1.9Deep bench#451–747−4.2
Every player on the 2026–27 preseason board, grouped by rank. Hover a dot to see who it is; the bar and the number at the right are the tier’s median.

LOKI against the predictive field

TestLOKIEPMDARKOLEBRON
Team win % explained (R²).70.65.62.59
…teams that reshuffled their roster.68.61.58.58
Game margins explained (R²).188.180.174.162
…games with reshuffled rosters.230.213.202.196

Preseason ratings predict the whole season, 2020–21 to 2025–26 (7,194 games, 180 team-seasons). R² is the share of variation the ratings explain.

Four releases a season

Each release uses only the games played before its date. The preseason board is frozen and hashed before opening night, then scored as the season is played.

OctNovDecJanFebMarAprMayJun
Regular seasonPlayoffsTodayR1R2R3R4
  1. R1 · Oct 19Preseason

    The whole board, frozen and hashed before tip-off. Scored live all season.

    Published
  2. R2 · ~Dec 13 (est.)Quarter mark

    First Sunday after the league reaches 375 games, 25 per team.

    Coming
  3. R3 · Feb 14Midseason

    First Sunday after the Feb 11 trade deadline: new rosters.

    Coming
  4. R4 · Apr 11Playoffs

    End of the regular season: the board heading into the playoffs.

    Coming

Current release: R1, 2026–27. SHA-256 a1f9bad844c3855d51703530d67102ed76c57eac80902115566d1716edd73bc9

LUMA Basketball ResearchLOKI one-pager · ARC one-pager · About · Changelog · Query terminalcourt-share.com/luma